@darkstar42
I set up an online entry point for fluency-loop so anyone visiting the repo can try the teaching skill directly, without cloning, installing Handy, or setting up the Python/PyTorch stack.
The design choice that stands out here is measuring spontaneous speech rather than drill accuracy — treating everyday Handy dictation as real practice data and tracking flow (speech rate, pause patterns, length of unbroken runs) instead of error counts. That distinction between fluency strength and storage strength, and the way the analyzer mines ambient dictation alongside prompted exercises, is a genuinely different bet from the usual language-learning feedback loop.
The tradeoff is that the setup path filters out almost everyone who would benefit from trying it: Claude Code + Handy + ffmpeg + Python 3.12 + PyTorch + Allosaurus is a serious stack before a single lesson appears. The online entry point gives visitors a way to run the teaching skill and see what a session produces before committing to that install. If someone lands on the README curious about the speak-measure-lesson loop, they can try it immediately — or forward the link to a friend learning a language — and come back for the full local setup once they see the output is worth it.
Every session through the entry point leaves a usage record, so you get signal on what language people pick, what missions they set, and where they get stuck — useful for sharpening SKILL.md and the onboarding flow.
Happy to send over the usage-record review link once sessions start coming through. Feel free to close this if it's not relevant.
shesonglin@tinkerland.ai
@darkstar42
I set up an online entry point for fluency-loop so anyone visiting the repo can try the teaching skill directly, without cloning, installing Handy, or setting up the Python/PyTorch stack.
The design choice that stands out here is measuring spontaneous speech rather than drill accuracy — treating everyday Handy dictation as real practice data and tracking flow (speech rate, pause patterns, length of unbroken runs) instead of error counts. That distinction between fluency strength and storage strength, and the way the analyzer mines ambient dictation alongside prompted exercises, is a genuinely different bet from the usual language-learning feedback loop.
The tradeoff is that the setup path filters out almost everyone who would benefit from trying it: Claude Code + Handy + ffmpeg + Python 3.12 + PyTorch + Allosaurus is a serious stack before a single lesson appears. The online entry point gives visitors a way to run the teaching skill and see what a session produces before committing to that install. If someone lands on the README curious about the speak-measure-lesson loop, they can try it immediately — or forward the link to a friend learning a language — and come back for the full local setup once they see the output is worth it.
Every session through the entry point leaves a usage record, so you get signal on what language people pick, what missions they set, and where they get stuck — useful for sharpening SKILL.md and the onboarding flow.
Happy to send over the usage-record review link once sessions start coming through. Feel free to close this if it's not relevant.
shesonglin@tinkerland.ai